perf: absorb column widths instead of re-measuring every row

[hellices]

Problem

After #208 a repaint costs only what changed, but the first build of a large
view still froze the UI for seconds — and so did every path that has to rebuild
rather than diff: switching kind, changing the sort, applying a custom view, and
the reorder fallback (a row inserted mid-table, since add_row can only append).

Splitting a 50,000-row cold build by wall clock showed the freeze is mostly not
our cell construction and not add_row:

phase	time
show() itself (build cells + add_row loop)	~1.3 s
Textual settling afterwards	~3.5 s

That settling is DataTable._update_dimensions(), which runs from on_idle over
every newly added row and, for each one, rebuilds the row's renderables through
default_cell_formatter and calls rich.measure() on every cell — 700,000
renderable constructions and 700,000 measurements, to compute fourteen integers,
from cells we had just finished building.

Approach

ResourceTable already knows every cell it emitted (_emitted, added in #208)
and already has _cell_width() — a cheap width function written to match
DataTable's own measurement, used today by the in-place diff.

1. _absorb_widths() folds the emitted cells straight into
   column.content_width, on both seeding paths (the rebuild path, and the rows
   the in-place diff appends).
2. _update_dimensions() is overridden to hand super() only the rows that
   were not absorbed. The superclass still measures anything added outside
   show() — including a row added after the absorption but before the idle pass
   — and still recomputes virtual_size exactly as before.
3. A cheap cannot-grow guard skips almost every measurement: an ASCII
   string's display width is exactly its length, and both markup parsing and the
   newline truncation _cell_width() performs can only shorten it. So
   len(raw) <= width and raw.isascii() proves the cell cannot widen its column,
   using two C-level calls and no measurement.

Failure mode if a future Textual renames the hook: the override simply stops
being called and the widget degrades to today's slower-but-correct measuring
pass.

Results

Cold show() + settle, same fixture, median of 3:

rows	before	after
1,000	142.5 ms	111.3 ms	1.3x
10,000	1010.4 ms	387.6 ms	2.6x
50,000	5017.1 ms	1507.0 ms	3.3x

The width absorption itself costs ~50 ms at 50,000 rows.

Tests

New tests/ui/test_table_column_widths.py:

• test_seeding_does_not_rebuild_every_row_to_measure_it — the RED test.
  Without the change 400 of 400 rows are rebuilt for measurement; with it only
  the handful painted on screen are.
• test_absorbed_widths_match_what_textual_would_measure — compares the
  resulting content_width against a control DataTable seeded with the same
  cells, over plain, long, full-width (CJK) and markup-bearing values. Identical.
• test_rows_added_outside_show_are_still_measured
• test_unabsorbed_row_added_before_the_dimension_pass_is_measured — guards the
  case where absorption and a direct add_row land in the same idle window.
• test_row_appearing_in_place_widens_its_column — a newly appeared long name
  still grows the column through the in-place diff.

Verification

make check on this branch: ruff ✅, ruff format ✅, mypy --strict ✅,
4081 passed / 21 skipped, tach ✅.

Not in scope

• Batch row insertion bypassing DataTable.add_row's per-row overhead (measured
  at ~2.8x on that phase, worth ~0.3 s of the remaining 1.5 s, but reaches much
  deeper into Textual internals).
• Chunked seeding across event-loop turns, which hides the remaining cost rather
  than removing it.

Fixes #210

Live-cluster validation

The numbers above are from synthetic summaries, which isolate the render path
but have uniform, all-ASCII cell content. To check the optimization holds on
real data, the same measurement was run against a live AKS cluster (v1.35.6)
holding 5,917 real pods across 6 namespaces, pulled through korvid's own
KubeClient — real manifests, real name/namespace length distribution, and two
custom columns (#45) sourced from annotations carrying Korean text, so the
non-ASCII path is genuinely exercised.

	value
pods	5,917
cells rendered	94,672 (89.4% ASCII, 10.6% full-width)
cold show() + settle, before	884.3 ms (median of 5)
cold show() + settle, after	314.5 ms (median of 5)
speed-up	2.8x
column widths identical to Textual's own measurement	✅

Widths came out byte-identical on both branches and matched the control
DataTable, including the two full-width columns (measured 30 and 45 cells wide
against far fewer characters) — so the isascii() fast path does not silently
under-measure CJK content.

[Copilot]

Pull request overview

Optimizes ResourceTable startup and rebuilds by reusing known cell widths instead of letting Textual remeasure every row.

Changes:

• Absorbs emitted-cell widths during rebuilds and appends.
• Filters already-measured rows from Textual’s dimension pass.
• Adds width-equivalence and fallback-path tests.

Reviewed changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated 1 comment.

File	Description
src/korvid/ui/widgets/resource_table.py	Implements column-width absorption.
tests/ui/test_table_column_widths.py	Tests performance invariant and width correctness.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[Copilot]

Pull request overview

Copilot reviewed 2 out of 2 changed files in this pull request and generated no new comments.

Suppressed comments (1)

src/korvid/ui/widgets/resource_table.py:628

• The required fail-safe for Textual's row-label/auto-height branches is missing: every emitted key is excluded without checking its Row metadata. Issue #210 explicitly requires the fast path to be disabled if the current height-1/unlabelled invariant changes; otherwise a future height=None row would keep height 0, or a labelled row would never update the label column. Only exclude rows that are confirmed height 1, non-auto-height, and unlabelled.

        absorbed = self._emitted
        super()._update_dimensions([key for key in new_rows if key.value not in absorbed])